Between an NPU and a TPU
We design low-power AI chips that pair a TPU-style matrix engine with programmable cores, built on mature 10 nm and larger processes to keep costs down. Cameras, wearables and machines get to make decisions on the device instead of waiting on the cloud.
Move your mouse over the city
The chip
Most AI on a device is matrix math, so we built a matrix engine for it and kept just enough programmability for everything else. The memory and sensor connections sit on the same chip, because moving data is where most of the power in AI hardware goes.
A TPU-style matrix engine does the heavy math. Small programmable cores handle the layers that don't fit it, so new kinds of models still run.
The model lives right next to the cores that use it. Fetching data from somewhere else costs more power than the math itself, so we avoid it.
Cameras and microphones plug straight into the chip, so a frame from a camera gets processed without passing through a separate main processor.
Bring a model you've already trained. Our tools convert it to run on the chip and tell you early on if something won't fit.
We make our chips on 10 nm and larger processes. They cost a fraction of the newest ones to produce and are easier to get factory time on. The power savings come from the design, not the process.
or email tanzim@ostolabs.com